TL;DR
This systematic review explores the burgeoning field of Artificial Intelligence (AI) in Obstetrics and Gynecology (OB/GYN) ultrasound imaging, highlighting its potential to revolutionize diagnostics and patient care. The paper synthesizes 189 studies, revealing a vast landscape of AI applications from fetal biometry to gynecological mass detection, with a particular emphasis on obstetrics.
What Problem Does This Paper Address?
Ultrasound (US) is a cornerstone of OB/GYN imaging due to its non-invasiveness and accessibility. However, traditional US imaging faces challenges: it's time-consuming, requires extensive specialized training, and is prone to inter- and intra-observer variability. Factors like image quality, patient characteristics (e.g., obesity), and fetal positioning can further complicate accurate interpretation. This leads to potential delays in diagnosis, increased workload for clinicians, and a need for more efficient and standardized approaches. AI offers a promising solution to automate tasks, improve accuracy, and enhance the overall efficiency of OB/GYN ultrasound examinations.
Methods
The authors conducted a systematic literature review, adhering to the PRISMA statement. They searched PubMed and the Cochrane Library for articles on AI applications in OB/GYN ultrasound. The search query included terms like "artificial intelligence," "deep learning," and "machine learning" combined with "ultrasound" and "obstetrics" or "gynecology." Articles were screened based on the PICOS framework, and full-text copies were analyzed. The included literature was categorized into obstetrics (148 articles) and gynecology (41 articles) based on their research topics. The review focuses on identifying the benefits and limitations of AI in OB/GYN US imaging, rather than assessing the effectiveness of specific AI technologies.
Datasets
Not explicitly detailed as distinct datasets for training or evaluation. The review synthesizes findings from 189 published research articles, implying that the datasets used were those collected and analyzed within those individual studies.
Key Insights and Ideas
- "5D Ultrasound" as a Metaphor: The paper uses "5D Ultrasound" to describe the enhanced dimension AI brings to traditional 2D/3D/4D US, encompassing image enhancement, automated calculations, and improved workflow.
- Obstetrics Dominance: The field of obstetrics has seen significantly more research into AI applications in US imaging compared to gynecology, likely due to the critical and frequent use of US in pregnancy monitoring.
- AI as an Assistant, Not a Replacement: A recurring theme is that AI is intended to augment, not replace, human expertise. It aims to reduce workload and improve accuracy, but clinical judgment remains paramount.
- Emerging vs. Established Fields: While fetal biometry is a well-established area for AI application, the review highlights emerging and experimental fields, such as AI for placental analysis and fetal neurosonography, suggesting avenues for future research.
- Ethical Considerations: The paper touches upon the ethical context of AI in healthcare, emphasizing the need for trustworthy AI and addressing the "black box problem" through explainable AI.
Key Results
The review identified 189 articles published between 1994 and 2023.
- Obstetrics Applications (148 articles):
- Fetal Biometry: The most common application, focusing on automated measurements of fetal head circumference, abdominal circumference, and femur length.
- Fetal Echocardiography: AI is used for automated acquisition of standard views and detection of congenital heart diseases.
- Fetal Neurosonography: Applications include automated detection of brain structures and malformations, and GA estimation.
- Other areas: Fetal face analysis, placental assessment, fetal malformation detection (first and second trimester), gestational age prediction, and workflow analysis.
- Gynecology Applications (41 articles):
- Adnexal Masses: AI is employed for automated discrimination between benign and malignant tumors.
- Breast Masses: Focus on detection, classification, and reducing unnecessary biopsies.
- Endometrium: AI assists in assessing endometrial thickness, texture, and identifying abnormalities.
- Pelvic Floor: Applications include assessing pelvic floor dysfunction and diagnosing prolapse.
- Other areas: Endometriosis, uterine fibroids, premature ovarian failure, follicle tracking, and ectopic pregnancy.
Limitations and Future Work
- Limitations:
- Expert Dependency: Many AI models still require expert input for image acquisition, ROI selection, and interpretation.
- Data Heterogeneity: Studies often suffer from small sample sizes, single-center data, or lack of external validation.
- Generalizability: AI models can struggle with subtle differences, variations in US probes/modalities, and different GA ranges.
- Real-time Application: The need for real-time application is often unmet, particularly for intrapartum scenarios.
- Pathological Cases: Many models are trained on healthy subjects, limiting their performance on pathological cases.
- Technical Focus: A significant portion of the literature comes from engineering, with a focus on technical aspects rather than clinical applicability.
- Future Work:
- Explainable AI: Developing AI that clinicians can trust and understand.
- Robustness and Generalizability: Training models on diverse datasets and validating them across different settings.
- Real-time Performance: Improving computational efficiency for real-time applications.
- Clinical Validation: Conducting more prospective studies to demonstrate clinical utility and impact.
- Emerging Fields: Further research into areas like placental health, fetal neurosonography, and AI for rare gynecological conditions.
- Interdisciplinary Collaboration: Bridging the gap between AI specialists and clinicians for better integration and trust.
Summary updated